This thesis focuses primarily on model and estimation uncertainty in empirical finance. We develop a “model portfolio” approach and a “global minimum variance” weighting scheme in a novel bias/variance trade-off framework. The thesis contains four essays.
The first essay develops the general research method. This essay shows that the “model portfolio” approach exhibits superior performance when the model space is incomplete and produces performance closest to a true model when the model space is complete. Using our modeling framework, the existing optimal weighting schemes can be unified.
The following chapters, which include the second, third and fourth essays of this thesis, apply our developed methodology to three distinct fields as follows: equilibrium asset pricing models, technical analysis and capital structure. At the same time, each essay also contributes to the specific area in which the application is implemented.
The second essay examines the performance of the “model portfolio” approach using equilibrium asset pricing models. This essay suggests that the “model portfolio” approach improves the single model and dynamic model selection method in out-of-sample forecasting and provides an economic explanation for this superior performance. The reason for the improved performance arises from the fact that single-model performance varies over time. A dominant model in one economic state may be inferior in other states. Given ex ante structural break uncertainty, a combined model helps hedge and diversify modeling uncertainty in the formation of forward-looking expectations.
The third essay illustrates the complementary role of three forecasting models, namely, the simple linear autoregressive model, the parametric non-linear technical moving average trading rule and the parametric cross-sectional momentum model. The model- encompassing tests show that all contain marginal information that can contribute to the “model pool”. Although all of the models are based on past market data, these three models interpret information in different ways and thus the combination of these models provides improved out-of-sample forecasting performance.
The fourth essay shifts from asset-pricing model uncertainty to corporate finance dynamic panel model estimation uncertainty problems. The essay illustrates our points in one concrete field: the speed of adjustment of corporate leverage in a partial adjustment framework. However, the method can be generalized to numerous other similar corporate settings.